OpenAI throttles GPT-5.6 rollout following U.S. government intervention
Table of Contents
A ‘Trusted’ Circle of Access
OpenAI is not releasing its latest suite of frontier models to the general public. Instead, the company has restricted the rollout of the GPT-5.6 lineup to a small group of “trusted partners,” a move OpenAI confirms was made at the behest of the U.S. government.
The new lineup introduces a tiered architecture designed to balance raw power with operational efficiency. At the top is Sol, the company’s new flagship model designed for complex reasoning. Below it sits Terra, intended for general everyday use, and Luna, a lightweight, low-cost option optimized for speed. While Sol represents the peak of OpenAI’s current capabilities, the Trump administration has effectively placed a hold on the public distribution of all three models.
According to OpenAI, the current preview is limited to specific partners whose identities have been shared with government officials. This restricted access is part of a growing trend of federal intervention in the deployment of “frontier AI,” following a series of aggressive moves by the administration to prevent advanced capabilities from falling into the hands of foreign adversaries.
The New ‘Involuntary Licensing’ Regime
This intervention follows a similar pattern seen with Anthropic. After the release of its Fable 5 model, the administration ordered the company to block access for foreign nationals, a request that ultimately led Anthropic to pull the model from public access entirely. This has sparked an intense debate among policy experts regarding the extent of executive power over private software releases.
Dean Ball, a former White House AI adviser who is transitioning into a role at OpenAI, suggests that the current environment has evolved into a de facto “involuntary licensing regime.” Under a recent executive order, certain AI firms are asked to voluntarily submit their most advanced models for government review up to 30 days before they hit the market. Ball argues that without clearly defined, objective safety standards, this process risks creating indefinite launch delays.
The stakes extend beyond corporate timelines. Ball warns that these restrictions could inadvertently assist China in the global AI race and jeopardize the massive capital investments currently flowing into AI infrastructure, as developers face uncertainty over when their products can actually be deployed.
The Technical Edge: Sol’s ‘Ultra’ Mode
Despite the rollout friction, the technical specifications of GPT-5.6 Sol suggest a significant leap in agentic capabilities. OpenAI claims the model excels in coding, biology, and cybersecurity, introducing a “max” reasoning mode and a high-tier “ultra” mode. The latter utilizes coordinated subagents to dismantle and solve highly complex tasks—a feature that provides immense power but significantly increases token consumption.
In early benchmarks, OpenAI asserts that Sol outperforms Anthropic’s Claude Mythos 5 in coding workflows. Furthermore, Sol is reportedly competitive with the Mythos preview while utilizing only one-third of the output tokens, suggesting a major improvement in efficiency and density.
To address safety concerns, OpenAI has moved away from the “classifier” approach that hindered Anthropic’s Fable 5. Rather than using a separate filter that routes risky prompts to older models—often resulting in false positives—OpenAI has integrated safety guardrails directly into Sol’s core behavior. The model is specifically optimized for defensive cybersecurity, designed to help users protect systems rather than provide instructions for offensive exploits.
Pricing and Availability
While the government-mandated restriction is currently in place, OpenAI describes this as a “short-term step.” The company is reportedly working with the administration to establish a more repeatable process for future releases and a new framework for cybersecurity.
When GPT-5.6 eventually reaches ChatGPT, Codex, and API users, it will follow a tiered pricing structure. Sol is positioned as the premium option at $5 per million input tokens and $30 per million output tokens. Terra is priced at half that rate, while Luna remains the most accessible at $1 per million input and $6 per million output tokens. To lower costs for developers, OpenAI has also implemented improved prompt caching to make repetitive queries more affordable.
